Ultrasound Thyroid Nodule Segmentation Based On Multi-branch and Color Space Volume

Ultrasonic segmentation of thyroid nodules is an indispensable part of computer aided system and diagnosis of thyroid diseases. Due to the fact that ultrasound images have the characteristics of asymmetric irradiation, uneven appearance, low signal-to-noise ratio, low contrast and blurred boundary, thyroid nodule segmentation is a challenging task. In order to improve the precision of thyroid nodule segmentation, this paper presents an Ultrasound thyroid nodule segmentation algorithm based on multi-branch and color space volume. In this method, an image pyramid is constructed to compensate for the spatial information loss by convolutional and pooling operations. We use pretrained ResNet block as the Feature Encoder Module, a multi-branch convolutional fusion (MBCF) block is proposed to increase the receptive field and capture high-level spatial and semantic information. To further improve the segmentation performance, the color space volume transformation is used to highlight the contrast between the target area and the background in the data preprocessing part. Experimental results on an Ultrasound thyroid nodule segmentation dataset show that the proposed method performs well over the state-of-the-art methods.